Age standardization using reweighting in complex survey data and its application in regression

Abstract

Age standardization is often used in analyzing complex survey data to compare subgroups with different age distributions. Conventionally, age-standardized estimates are derived as weighted averages of age-specific estimates applying weights based on a standard population of choice. Age adjustment, to be distinguished from age standardization, is compatible with widely-used, regression-based techniques for assessing time trends. Those techniques are much harder to use in conjunction with the conventional age standardization method. A reweighting-based alternative method, which can readily accommodate regression-based time trend assessment, is described in this presentation. The alternative involves modification of sampling weights that force age composition for each subgroup to match that of the standard population. To illustrate an application of the alternative, time trend assessments of erythrocyte metal measurements by race/Hispanic origin subgroup in the National Health and Examination Nutrition Survey (NHANES) 1999-2016 are used. The estimand of interest is age-standardized geometric mean derived from Tobit regression that treats measurements below the lower limit of detection as left-censored. When ordinary least square-type regression was used in place of Tobit regression, the reweighting-based method produced age-standardized estimates that exactly matched those obtained using the conventional method, as expected mathematically. When Tobit regression was used, the conventional and alternative methods produced slightly different age-standardized geometric mean estimates for erythrocyte concentration of lead, cadmium, and mercury by sex, race/Hispanic origin, and survey cycle with the difference within ± 5% in 193 out of 198 metal-subgroup combinations. The reason for the difference was investigated to inform potential method improvement for reducing the difference. Annual percentage changes and p-values from orthogonal contrast-based tests for linear, quadratic, and other higher order trends can be produced easily in conjunction with the reweighting-based age standardization.